The maths of the boom in artificial intelligence is starting to undermine its foundations, with the implosion in the share price of a leading neocloud last week, and soaring bond yields, highlighting the most vulnerable point in a fragile financial edifice.
Last week the share price of neocloud CoreWeave plummeted 15 per cent and the yield on its bonds – bonds it issued with a yield of 9.675 per cent in June – rocketed to 15 per cent, trading at less than 80 per cent of their face value.
CoreWeave wasn’t alone. The share prices and bonds of other neoclouds – companies that are built specifically to provide computing power to train AI models – were also ravaged.
Earlier this month, data centre operator Firmus yanked what would have been Australia’s second-largest initial public offering after investors balked at the $44 billion valuation and the very pricey capitalisation of revenues and earnings from projects that are still on drawing boards.
It probably wasn’t a coincidence that the plunge in the value of CoreWeave’s shares and bonds came in the wake of the Firmus debacle and accelerated as last week’s data showed the US inflation rate is also accelerating, presaging higher US interest rates.
The business model for the neoclouds – indeed for most of the AI industry – is novel.
They commit to building or leasing data centres and installing the massive computing power necessary for AI, using contracts to rent out that capacity to AI frontier labs and the hyperscalers as the backing for the equity and debt that they raise to fund the projects.
In CoreWeave’s case – and this is a common and much questioned feature of the AI sector – it has committed to buying billions of dollars of high-end graphics processing units (GPUs) or chips, from Nvidia. In turn, Nvidia helped to buttress CoreWeave’s finances by investing $US2 billion in the company and providing a $US6.3 billion commitment to buy any unsold capacity through to 2032.
In effect, Nvidia has helped fund the purchase of its own chips, a feature often described as “circular” financing and one rife throughout the sector.
CoreWeave has a not dissimilar arrangement with OpenAI, which has committed to pay more than $US22 billion for “compute” while taking up a $US350 million equity stake in the company.
Those contracts with major AI players – there’s another with Meta – have enabled CoreWeave to raise the equity and debt to fund its data centres rollout. It has more than $US40 billion of debt and leases.
The challenge for not just CoreWeave but the sector generally, is that the commitments to landlords are long term, but the subleases to customers are for short-term timeframes. Adding to the risk is the rate of development of the GPUs – the chips they are installing at increasingly exorbitant cost may be obsolete by the time their centres are operational.
Moreover, thanks to a community backlash against data centres (no-one wants to live next to them and they can consume vast amounts of energy and water) and the inflation rate, the facilities are taking longer than expected to be built and are more expensive than anticipated.
As with AI more broadly, there is a significant mismatch between the timing of the capital expenditures and revenue growth, which is widening because of the pace at which the sector is growing. That yawning and ever-expanding gap has to be funded by ever-increasing amounts of equity and debt.
Last month the global technology research and advisory firm Gartner released an analysis that concluded that AI companies were unlikely to deliver enough revenue to cover the expected capital expenditure on IT infrastructure between now and 2030.
It said AI providers would spend nearly $US1 trillion on infrastructure this year and were forecast to spend another $US6 trillion by 2030. Its research, however, indicated that enterprise and consumer revenues would be insufficient to support profitable returns on the spending.
The operators would need to generate $US8.5 trillion of revenue to cover the capital outlays and funding costs while generating a 12 per cent return on capital (which would be the minimum required, given the riskiness, uncertainty and protracted timelines of AI investments).
The total revenues Gartner forecasts for AI data centres through to 2032, however, only amount to $US3.6 trillion. It said more than 80 per cent of the neocloud operators would cease operating independently by 2030.
One of the oddities of the entire AI boom is how undiscriminating equity investors and, until very recently, debt investors have been.
They’ve been willing to supply vast amounts of cheap capital to all things AI-related in what have been blind acts of faith. It is anyone’s guess what the longtime returns on the investments in models, chips and data centres will eventually be.
As we know from the experience of the 19th century railways boom, and the dot-com boom of the 1990s, AI could transform the world as we know it and still be a poor – even disastrous – investment today.
The experience of Firmus, CoreWeave and their peers – and analyses like Gartner’s – suggests scepticism is starting to surface, and risk is starting to be priced in to equities and debt. Even Nvidia’s share price has fallen by more than 10 per cent in the past fortnight.
CoreWeave’s bond yields scream distress. Its credit default swaps – it now costs almost US1.2 million to ensure $US10 million of its debt – are pricing in a more than 60 per cent chance that it will default within the next five years.
The company’s June quarter results highlight the challenge it confronts. It generated revenue of $US2.6 billion, had capital expenditures of $US9.4 billion and lost $US626 million. Its free cash flows were negative to the tune of $US5.74 billion.
Both CoreWeave and Nvidia have defended the group’s financials, with CoreWeave’s chief executive Mike Intrator saying the markets are acting “irrationally” and he had confidence in the business model.
That confidence, however, revolves around the markets’ willingness to continue to supply the ever-swelling amounts of capital that fills in the widening gap between the company’s cash generation and its spending commitments.
With the bonds of neoclouds trading below their face value, and their shares imploding, that willingness is going to be tested, along with the sturdiness of the entire ASI sector’s foundations.
Rising interest costs are changing the economics of AI, increasing the costs and deferring the moment of hoped-for profitability further into the future, and, by devaluing the net present value of the hoped-for future cash flows, lowering the value and appeal of their equity today.
If the supply of that capital, however expensive, were to be throttled, however, it would be extremely destructive – fatal for many – for all companies with exposures to AI. The probable exception is the stronger hyperscalers, which have their own legacy cash flows to fall back on.
The Market Recap newsletter is a wrap of the day’s trading. Get it each weekday afternoon.
From our partners
About the Author
